Convolution over Hierarchical Syntactic and Lexical Graphs for Aspect Level Sentiment Analysis

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 3540–3549, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 3540 Convolution over Hierarchical Syntactic and Lexical Graphs for Aspect Level Sentiment Analysis Mi Zhang, Tieyun Qian∗ School of Computer Science, Wuhan University, China {mizhanggd, qty}@whu.edu.cn Abstract The state-of-the-art methods in aspect-level sentiment classification have leveraged the graph based models to incorporate the syntac- tic structure of a sentence. While being ef- fective, these methods ignore the corpus level word co-occurrence information, which reflect the collocations in linguistics like “nothing special”. Moreover, they do not distinguish the different types of syntactic dependency, e.g., a nominal subject relation “food-was” is treated equally as an adjectival complement relation “was-okay” in “food was okay”. To tackle the above two limitations, we pro- pose a novel architecture which convolutes over hierarchical syntactic and lexical graphs. Specifically, we employ a global lexical graph to encode the corpus level word co-occurrence information. Moreover, we build a concept hierarchy on both the syntactic and lexical graphs for differentiating various types of de- pendency relations or lexical word pairs. Fi- nally, we design a bi-level interactive graph convolution network to fully exploit these two graphs. Extensive experiments on five bench- mark datasets show that our method achieves the state-of-the-art performance. 1 Introduction Aspect-level sentiment classification (ASC) (Hu and Liu, 2004) aims to determine the sentiment polarity (i.e., positive, negative, neutral) of the as- pect(s) in a sentence. Take the review “great food but the service was dreadful” as an example. Given two aspect terms “food” and “service”, the goal is to infer the sentiment polarities for the aspect terms: positive for food and negative for service. ASC can provide fine-grained analysis of the users’ opinion towards the specific aspect and is fundamental to *Corresponding author. many natural language processing tasks. Conse- quently, it has aroused much research attention in recent years. Early studies on ASC (Mohammad et al., 2013; Jiang et al., 2011) mostly use machine learning algorithms to build sentiment classifier. Later, vari- ous neural network models (Dong et al., 2014; Vo and Zhang, 2015; Chen et al., 2017) are proposed for this task, including long short-term memory (LSTM) based (Wang et al., 2016), convolutional neural networks (CNN) based (Huang and Carley, 2018; Li et al., 2018), and memory based (Tang et al., 2016b) or hybrid methods (Xue and Li, 2018). These models represent the sentence as a word sequence and neglect the syntactic relations be- tween words, and thus it is hard for them to find the opinion words which are far away from the as- pect term. To solve this problem, several recent researches (Zhang et al., 2019; Huang and Carley, 2019; Sun et al., 2019) leverage the graph based models to incorporate the syntactic structure of a sentence, and have shown better performance than those without considering syntactic relations. Despite of their effectiveness, the seminal syn- tax based methods ignore the corpus level word co-occurrence information. Moreover, they do not distinguish the …

Content truncated to 500 words